def net2net_deeper_nononline(self):
s = deeper(self.conv1,
None,
bnorm=False,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv1 = s
s = deeper(self.conv2,
None,
bnorm=False,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv2 = s
s = deeper(self.conv3,
None,
bnorm=False,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv3 = s
def net2net_deeper(self):
s = deeper(self.conv1,
F.ReLU,
bnorm=False,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv1 = s
s = deeper(self.conv2,
F.ReLU,
bnorm=False,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv2 = s
s = deeper(self.conv3,
F.ReLU,
bnorm=False,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv3 = s
def net2net_deeper(self):
s = deeper(self.conv1,
nn.ReLU,
bnorm_flag=True,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv1 = s
s = deeper(self.conv2,
nn.ReLU,
bnorm_flag=True,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv2 = s
s = deeper(self.conv3,
nn.ReLU,
bnorm_flag=True,
weight_norm=args.weight_norm,
noise=args.noise)
self.conv3 = s
print(self)
def net2net_deeper_recursive(model):
"""
Apply deeper operator recursively any conv layer.
"""
for name, module in model._modules.items():
if isinstance(module, nn.Conv2d):
s = deeper(module, nn.ReLU, bnorm=False)
model._modules[name] = s
elif isinstance(module, nn.Sequential):
module = net2net_deeper_recursive(module)
model._modules[name] = module
return model
def test_deeper(self):
net = self._create_net()
inp = th.autograd.Variable(th.rand(32, 1, 28, 28))
net.eval()
out = net(inp)
s = deeper(net._modules['conv1'], nn.ReLU, bnorm_flag=True, weight_norm=False, noise=False)
net._modules['conv1'] = s
s2 = deeper(net._modules['conv2'], nn.ReLU, bnorm_flag=True, weight_norm=False, noise=False)
net._modules['conv2'] = s2
s3 = deeper(net._modules['fc1'], nn.ReLU, bnorm_flag=True, weight_norm=False, noise=False)
net._modules['fc1'] = s3
net.eval()
nout = net(inp)
assert th.abs((out - nout).sum().data)[0] < 1e-1
# test for 3D net
net = Net3D()
inp = th.autograd.Variable(th.rand(32, 1, 16, 28, 28))
net.eval()
out = net(inp)
s = deeper(net._modules['conv1'], nn.ReLU, bnorm_flag=False, weight_norm=False, noise=False)
net._modules['conv1'] = s
# s2 = deeper(net._modules['conv2'], nn.ReLU, bnorm_flag=False, weight_norm=False, noise=False)
# net._modules['conv2'] = s2
net.eval()
nout = net(inp)
assert th.abs((out - nout).sum().data)[0] < 1e-1, "New layer changes values by {}".format(th.abs(out - nout).sum().data[0])